Exploration-Lab/IL-PCSR-Models

IL-PCSR (Indian Legal — Precedent & Statute Retrieval)

Ensemble Model: A hybrid approach combining lexical features (BM25 5-gram) with semantic/distributional features (Para-GNN) with dynamic weighting between features, that is effective for both legal statute as well as prior case retrieval.

Summary of Model files

We have 5 files for 3 different types of models:

• only_secs_model.bin, only_precs_model.bin — separate models for LSR and PCR, ft. independently
• multi_task_model.bin — single model for both LSR and PCR, ft. together in a multi-task setup
• pipeline_secs.bin, pipeline_precs.bin — separate models for LSR and PCR obtained via transfer learning (pipeline_secs.bin is obtained by LSR training on only_precs_model.bin, i.e., transfer PCR --> LSR, and vice versa)

All of these models have been trained with summaries of queries and precedents, and not full documents.

---

How to Use

All of the examples assume you have access (i.e., gate accepted). You need to use huggingface_hub to download the model contents to a local file, after which it can be loaded like any standard PyTorch model.

from huggingface_hub import hf_hub_download

# Download to local file
file_path = hf_hub_download(
    repo_id="Exploration-Lab/IL-PCSR-Models",
    filename="multitask_model.bin"
)

print("Model weights downloaded to:", file_path)

import torch

# Load the state dict in pytorch
trained_state_dict = torch.load(file_path, map_location=torch.device('cpu'))

---

Citation

@inproceedings{il-pcsr2025,
  title = "IL-PCSR: Legal Corpus for Prior Case and Statute Retrieval",
  author = "Paul, Shounak and Ghumare, Dhananjay and Goyal, Pawan and Ghosh, Saptarshi and Modi, Ashutosh"
  booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
  month = nov,
  year = "2025",
  address = "Suzhou, China",
  publisher = "Association for Computational Linguistics",
  note = "To Appear"
}